3-Tier Agent Architecture (No-Haiku)

MoAI-ADK v3.0 excludes Haiku from the routing model set and distributes work across a 3-tier structure keyed to task character. This design is grounded in empirical data from the DeepSWE leaderboard. This page explains why Haiku is excluded, how the 3 tiers are configured, and distinguishes design intent from implemented behavior.

The key finding from the DeepSWE leaderboard is that “weak model + high effort = enemy of availability.” A weaker model does not finish a long-horizon task more cheaply — it spends more steps and more output tokens failing to converge. At max effort, Sonnet 5 consumes 268 steps and 214k output tokens on the same task set that Opus 5 finishes in 99 steps.

Measurements below are from the leaderboard’s “All effort levels” view (113 tasks / 91 repos / 5 languages, mini-swe-agent harness). Because effort is reported per level, the tiers can be derived from the shape of each model’s cost/score curve rather than from a single operating point.

List price per MTok (in/out): Opus 5 $5/$25 · Sonnet 5 $2/$10 (introductory, through 2026-08-31, then $3/$15) · Fable 5 $10/$50.

Price inversion: Sonnet’s per-token price is below Opus, yet its per-task cost is higher at every comparable point — Opus 5 at low costs $1.66 and scores 58%, while Sonnet 5 at max costs $26.40 and scores 54%. The conventional wisdom that “running a cheaper model saves quota” does not hold for long-horizon agentic work, because completion efficiency, not unit price, sets the bill.

Under this data, including Haiku in routing would add step waste without adding capability. Sonnet is instead confined to single-shot, input-dominated work where multi-step completion failure does not apply.

Models and effort are assigned to 3 tiers based on task character.

flowchart TD
    T1["Tier 1 — Single-shot
Sonnet low
git mechanics · read-only search"]
    T2["Tier 2 — Agentic
Opus low/medium/high
spec · develop · audit · design · harness"]
    T3["Tier 3 — Peak
Opus max
develop · advisor (high profile only)"]

    T1 --> T2 --> T3

Work that completes in one pass and is dominated by input rather than by iteration. Multi-step completion failure — the effect that makes weaker models expensive — does not apply here, so Sonnet’s lower input price is the operative factor. Sonnet at low effort minimizes step count. Agents: manager-git, Explore. These two rows are fixed across all three profiles.

Every multi-turn row — planning, implementation, audit, design, harness generation, documentation, E2E. Opus carries all of them, because Opus at low already outscores Sonnet at any effort while costing less per task. The profile selects where on the Opus effort ladder each row sits: low in the economical column, medium in the default column, high in the quality column. Agents: manager-spec, manager-develop, plan-auditor, sync-auditor, manager-design, builder-harness, manager-docs, e2e-tester.

max effort is confined to the two rarest-invocation rows in the high profile only: manager-develop and super-advisor. Above medium the marginal cost per point rises sharply ($0.15/point for low→medium vs $0.70/point for medium→high), so peak effort is spent only where a single decision carries disproportionate downstream cost. xhigh is used nowhere — on Opus it matches high on score at 49% higher cost.

Four conclusions drawn from the per-effort measurements:

1. Opus 5 Pareto-dominates Sonnet 5 at every effort. Opus at low (58%, $1.66) beats all five Sonnet points on both axes, including Sonnet at max (54%, $26.40). The routing thesis “send busy agents to the cheaper model” is falsified for long-horizon agentic work.
2. The cause is completion efficiency, not price. Sonnet spends roughly 2.7× the steps to finish the same task set. The extra steps and output tokens, not the per-token rate, are what make the task expensive.
3. xhigh is a net loss on Opus. high and xhigh both score 73%, but xhigh costs 49% more and takes 22% more steps. The same flat top appears on Fable. Effort past the knee buys tokens, not points.
4. medium is the knee. Marginal cost per point: low→medium $0.15, medium→high $0.70 (4.7×), xhigh→max $2.77 (18.6×). The default profile anchors manager-develop at medium for this reason.

Limitation note: the benchmark measures coding agents. Documentation authoring, audit judgment, and SPEC authoring quality are not directly measured, so those row placements rest on a similarity inference to multi-turn agentic work rather than on observation. Confidence intervals also matter: medium (69%±1) and high (73%±2) do not overlap, but max (74%±4) overlaps high — which is why max is confined to two rarely-invoked cells. Every default is reversible per-agent via llm.agent_overrides.

On Fable 5: Fable is dominated on the coding axis at every effort — Fable at high (69%, $9.18) matches Opus at medium (69%, $3.29) for nearly triple the cost — so it appears in no matrix cell. It remains a valid value in the model enum and stays wired as the GLM backend’s Fable slot; only the defaults changed.

REQ-DA-061 honesty distinction: The content on this page must clearly distinguish design-stage from implemented behavior.

Design stage (.moai/reports/agent-architecture-redesign-v2-20260709.html) — the v2 architecture design intent. Presents the 3-tier model policy principles and DeepSWE rationale.

Implemented behavior — a single profile matrix performs the actual routing. The active profile (high/medium/low) selects one column of the matrix, and the resolver determines each agent’s {model, effort} and injects the model as a runtime argument at spawn time. For the detailed matrix, see the Profile Matrix page.

Readers must be able to distinguish design intent (the DeepSWE rationale on this page) from implemented behavior (the single profile matrix).

The 3-tier architecture is the substrate for harness self-evolution. For the evolution loop (observation → reflection → promotion) to be effective, routing decisions in the observation phase must go to the right model at the right effort. For details on self-evolution, see the Harness Self-Evolution page.

• Profile Matrix — the single 3-column per-agent profile matrix (11 agents × 3 profiles = 33 cells)
• Tokenomics Overview — Layer B routing of the 4-layer tokenomics structure